The diagnostic accuracy/efficacy of MRI in differentiating hepatic hemangiomas from metastatic colorectal/breast carcinoma: A multiple reader ROC analysis using a jackknife technique
Purpose: Our purpose was to determine the diagnostic accuracy efficacy of a simple MR technique in differentiating hepatic hemangiomas from colorectal or breast metastases using a multiple reader method. Method: Thirty-seven cases with confirmed hepatic hemangiomas and 115 with confirmed hepatic metastases (colon primary, n = 86; breast primary, n = 29) evaluated with MRI at 1.5 T were retrospectively collected. A single lesion in a single slice from each patient was randomly selected; the images were masked and then were interpreted in random order by five separate readers blinded to the diagnosis using a five point diagnostic scale (from definite hemangioma to definite metastasis). Morphologic characteristics of lesion margin, signal intensity relative tot other structures, and internal architecture (homogeneous versus heterogeneous) were also assessed independently of the five point diagnostic scale. Three of the readers had >8 years of experience, while the other two had 1 and 3 years. The diagnostic scale results were subjected to receiver operating characteristic (ROC) analysis using a jackknife method. κ-Statistics were applied to assess interreader agreement in the morphologic characteristics. A logistic regression model was used to determine which characteristics predicted pathology and reader diagnosis. Results: ROC analysis showed the average area under the curve over all readers was 0.91 (0.89-0.93 95% confidence interval) (p < 0.0001). An analysis of variance showed no significant difference between the areas under the curves of each reader (p = 0.6433). When the definite and probable categories for hemangioma and metastasis were combined, the sensitivity/specificity for the diagnosis of hemangioma ranged from 57 to 73%/91 to 97%. The positive/negative predictive value ranged from 72 to 84%/87 to 91%. For the morphologic assessment, there was significant agreement between the readers (p < 0.0001-0.0037). A sharp margin and lesion signal equal to or greater than CSF predicted the presence of a hemangioma (p = 0.0148 and p < 0.0001, respectively). A sharp margin, lesion signal equal to or greater than CSF, and a homogeneous internal architecture all predicted the reader diagnosis of definitely or probably hemangioma. Conclusion: For multiple readers, T2-weighted SE MRI alone is a very specific method for distinguishing hemangiomas from metastatic colon or breast carcinoma. Morphologic characteristics of a sharp margin and a high signal predict the presence of a hemangioma. Last, reader experience does not appear to have a significant effect on the specificity.
Lee, M-G; Baker, ME; Sostman, HD; Spritzer, CE; Paine, S; Paulson, EK; Keogan, MT
Volume / Issue
Start / End Page
International Standard Serial Number (ISSN)
Digital Object Identifier (DOI)